Background
The Radiological Society of North America (RSNA) receives more than 8000 abstracts yearly for scientific presentations, scientific posters, and scientific papers. Each abstract is assigned manually one of 16 top-level categories (e.g. "Breast Imaging") for workflow purposes. Additionally, each abstract receives a grade from 1-10 based on a variety of subjective factors such as style and perceived writing quality. Using machine learning to automate, at least partially, the categorization of abstract submissions can result in saving many hours of manual labor.
Methods
A total of 45527 RSNA abstract submissions from 2014 through 2019 were ingested, tokenized, and pre-processed with a standard natural language programming protocol. A bag-of-words (BOW) model was used as a baseline to evaluate two more sophisticated models, convolutional neural networks and recurrent neural networks, and also evaluate an ensemble model featuring all three neural networks.
Results
ensemble model was able to achieve 73% testing accuracy for classifying the 16 top-level categories, outperforming all other models. The top model for classifying abstract grade was also an ensemble model, achieving a mean average error (MAE) of 1.01.
Conclusion
While the baseline BOW model was the highest performing individual classifier, ensemble models that included state-of-the-art neural networks were able to outperform it. Our research shows that machine learning techniques can, to a reasonable degree of accuracy, predict both objective factors such as abstract category as well as subjective factors such as abstract grade. This work builds upon previous research involving using natural language processing on scientific abstracts to make useful inferences that address a meaningful problem.
Machine learning, natural language programming, and electronic health records: The next step in the artificial intelligence journey?
